On Thu, Jun 11, 2026 at 03:12:39PM +0800, [email protected] wrote: > From: Menglong Dong <[email protected]> > > For now, we use sk_busy_loop() in __xsk_sendmsg() to send the data in tx > ring. The sk_busy_loop() will poll on the target NAPI. However, for the > nic driver that support the tx napi, such as virtio-net, it can't schedule > the tx NAPI, but only the rx NAPI. If we enable the busy_poll for xsk and > use virtio-net, we can't send data, as the rx NAPI in virtio-net doesn't > handle the packet sending.
Am I reading this right that you decided to break busy-poll support for zero-copy drivers that happen to handle transmit side from Rx NAPI context in favor of supporting virtio-net? > > Fix this by introduce the sk_tx_busy_loop(), which will poll on the tx > NAPI if available. To get the tx NAPI from the napi_id, we add the > "tx_napi" field to napi_struct, which is ugly :/ > > Another choice is to call virtnet_xsk_xmit() in virtnet_poll() too. But > this a little contradict the design of tx NAPI. > > Menglong Dong (3): > net: busy-poll: introduce sk_tx_busy_loop() > virtio_net: initialize napi.tx_napi in virtnet_alloc_queues() > xsk: replace sk_busy_loop with sk_tx_busy_loop in __xsk_sendmsg() > > drivers/net/virtio_net.c | 1 + > include/linux/netdevice.h | 1 + > include/net/busy_poll.h | 41 ++++++++++++++++++++++++++++++++++++--- > net/core/dev.c | 23 +++++----------------- > net/xdp/xsk.c | 2 +- > 5 files changed, 46 insertions(+), 22 deletions(-) > > -- > 2.54.0 >
